ML Directors on Landmine Detection, Deadheading Trucks, and Media's Makers

Machine learning directors rarely get asked the interesting questions. The job description demands fluency in model architectures, data pipelines, and stakeholder management — but the people doing the work have a sharper view of what actually moves the needle. A new series from Inblix taps that group directly, pulling insights from directors across healthcare, logistics, media, and research.

Archi Mitra, a director at BuzzFeed, frames media’s ML challenge around a tension most industries never face. Privacy-first personalization means serving users whose long-term interests are stable but whose short-term interests are, in his words, “stochastic.” The hard part isn’t the recommendation engine — it’s respecting editorial voice and equitable coverage while doing it. Mitra also points to a failure mode he sees constantly: teams build ML systems without involving the editors, writers, and creators who give media its voice. Those “makers,” as he calls them, are the larynx of the business. Ignore them and the model optimizes for clicks while hollowing out the product.

Elsewhere in the series, one director does some back-of-napkin math on trucking logistics. Empty deadheading — trucks driving without cargo — happens roughly 20% of the time. Cutting that by a single percentage point, to 19%, would reduce carbon emissions by an amount equivalent to removing about 100,000 Americans from the road. The estimate comes from an ex-rocket scientist, so the napkin has some credibility. Another featured director works on ground-penetrating radar to detect buried landmines, a reminder that ML’s most consequential applications aren’t always the ones grabbing headlines.

The throughline across these interviews is less about model performance and more about constraints. In media, the constraint is editorial integrity. In trucking, it’s operational reality. In humanitarian demining, it’s a false negative that costs a limb or a life. Directors who thrive in these environments aren’t the ones chasing state-of-the-art benchmarks — they’re the ones who understand the system their model operates inside. That distinction gets lost in most AI coverage, which is exactly why this series exists.